Dice Question Streamline Icon: https://streamlinehq.com

Effect of mixing HRRR forecast and analysis data on 40 dBZ reflectivity performance

Determine whether training HRRRCast with a mixture of High-Resolution Rapid Refresh (HRRR) forecast fields and HRRR analysis fields improves composite reflectivity prediction skill at the 40 dBZ threshold, relative to training solely on HRRR analysis data.

Information Square Streamline Icon: https://streamlinehq.com

Background

HRRRCast is trained on HRRR analysis data to leverage the highest-quality fields available. Prior work noted that training on forecast data can yield sharper reflectivity at short lead times, but the authors emphasize that the analysis dataset is preferable for their longer (up to 48 h) lead times.

In the qualitative evaluation, the authors observe HRRRCast aligns more closely with HRRR analysis than the HRRR forecast, yet HRRR performs slightly better at the moderate 40 dBZ threshold. This motivates investigating whether mixing forecast and analysis data in training could enhance performance at 40 dBZ.

References

While the analysis dataset is of higher quality--especially relevant for HRRRCast, which supports forecast lead times up to 48 hours--it remains an open question whether incorporating a mix of forecast and analysis data could enhance performance, particularly at the 40 dBZ reflectivity threshold. We plan to explore this in future work.

— HRRRCast: a data-driven emulator for regional weather forecasting at convection allowing scales (2507.05658 - Abdi et al., 8 Jul 2025) in Section 3.1 (Qualitative evaluation)